search

sigsep / open-unmix-pytorch

Open-Unmix - Music Source Separation for PyTorch
https://sigsep.github.io/open-unmix/
MIT License
1.24k stars 181 forks source link

multichannel generalized Wiener filter question #125

Open nnanos opened 2 years ago

nnanos commented 2 years ago

Is it possible to train the Open-Unmix with a different Front-End (e.g. different parameters of the STFT ) for each source and then on the evaluation phase use the multichannel generalized Wiener filter?? I dont know the mechanics of the multichannel generalized Wiener filter but by inspection of the evaluation.py code you CANT provide it with different separator.json files (one for each Front End) . Thus I think that If I want to train and evaluate OPen-Unmix with different FEs this can be done only if :

1)Play arround with the code in order for the evaluation.py to accept different separator.json files. 2)Obviously dont use the multichannel generalized Wiener filter step.

Thank you for your time.

faroit commented 2 years ago

Good question, the time-frequency resolutions have to be the same. For learnable front-ends I'm not sure if anyone has tried this yet.